The Substrate
Measurement science embedded into the foundation
Most platforms collect data and hand it to a model. We built the measurement science directly into how data is structured, reconstructed, and modeled.
Data architecture
What goes in
The inputs that feed the model — individual-level data is optional. The model fuses aggregate and user-level signals, and estimates exposure where it cannot be directly measured.
Purchasing Data
→
Media Exposure
→
Demographic Data
→
Journey Model
→
Attribution ΔP
THE SCIENCE
A model of how consumers decide
Marketing data doesn't arrive complete. Journeys are fragmented. Exposure is estimated. Gaps are everywhere. Most platforms fill those gaps with assumptions. Ours fills them with a model — a discrete-choice framework built on the same economic science used to model real human decisions. That model doesn't just run on the data. It shapes how the data is reconstructed in the first place.
Journey model: Foundations
Econometric discrete-choice framework
Consumer-level measurement science
Consumer-level choice modeling
Links individual consumers' choices to stimuli — marketing and exogenous factors — across geo markets.
Rather than fitting a curve to sales history, the model estimates why a consumer chose your brand over a competitor — and what would change that decision.
Likelihood maximization
Maximizes coefficient likelihood given data — not prediction error — enabling credible scenario planning.
Optimized to estimate the true effect of each lever, not to replay history. That's what makes “what if” questions answerable.
User-level & aggregate data fusion
Exposure modeled to correct endogeneity and impute gaps; individual-level data is optional.
Corrects for the error where advertising looks more effective than it is — because brands naturally spend more when sales are already rising.
Consumer-level choice modeling
Prior-period results carried forward as priors, mimicking human learning and stabilizing outputs over time.
When data is missing or unreliable, the model fills gaps using everything it already knows — prior results, category behaviour, market context.
LOG-LIKELYHOOD FUNCTION
Σ log P(ȳₜ | x̄, {yᵢₜ}, θ) + ΣΣ log P(yᵢₜ | xᵢₜ, θ)
Aggregate LL + User-Level LL (discrete choice)
Attribution: Counterfactual Method
Simulated media removal at consumer level
Model-driven, not rule-based
Simulate media removal
For each consumer journey, remove a single channel and compute the resulting change in purchase probability.
Not credit allocation — a direct measurement of what each channel actually caused.
Causal delta attribution
Attribution = P(KPI=1) − P(KPI=1 | channel=0) — model-driven, not rule-based last-touch.
The drop in purchase probability when a channel is removed is its true incremental contribution.
Cross-channel interaction effects
The model captures the full journey sequence, not isolated touchpoints — handling complex channel interplay.
TV + social + search working together is modelled as a sequence, not three separate contributions that sum to 100%.
Unified multi-channel framework
Applied across paid social, video, display, and OOH within a single consistent model.
One model, one methodology — no stitching together outputs from separate platform measurement tools.
Observed journey
FB
→
YouTube
→
Billboard
→
Display
→
Purchase
Observed journey
FB
→
YouTube
→
Billboard
→
Display
→
ΔP = YouTube impact
Removing YouTube from the journey and measuring the drop in purchase probability gives YouTube's true incremental contribution — not a rule-based allocation.
Academic foundations
Built on 50 years of economic science
Our measurement framework is grounded in peer-reviewed econometric research.
Put the science to work for your marketing team.
01
McFadden (1974)
02
Berry, Levinsohn & Pakes (1995)
03
Manchanda, Rossi & Chintagunta (2003)
04
Bollinger, Cohen & Jiang (2013)
05
McCarthy & Oblander (2020)